Choosing an SDK¶
Most teams don't need to choose: the gateway accepts the OpenAI, Anthropic and Gemini request formats, and translates between them, so the SDK you already have can reach every provider your organisation has approved. Changing the base address and the key is the whole migration.
The short answer¶
| You have… | Use | Base address |
|---|---|---|
| Code on the OpenAI SDK, or a tool that says "OpenAI-compatible" | The OpenAI format | https://<your-gateway>/v1 |
| Code on the Anthropic SDK, Claude Code or the Claude Agent SDK | The Anthropic format | https://<your-gateway> |
| Code on the Google GenAI SDK or Google ADK | The Gemini format | https://<your-gateway>, API version v1beta |
| Nothing yet | The format of the provider whose own features matter most to you | as above |
Whichever you pick, the key is a Vidai key (starts vk_), and every call is checked against your policies, charged
to the right team and recorded in Request logs.
What crosses providers and what doesn't¶
| Calls | Reach |
|---|---|
Chat: /v1/chat/completions, /v1/messages, :generateContent and :streamGenerateContent |
Every provider: OpenAI, Azure OpenAI, Anthropic, Gemini, Vertex AI and Amazon Bedrock, translated both ways, with streaming and tool calls |
| Everything else (Responses, embeddings, images, audio, files, batches, cached contents…) | Only a provider that speaks the same format; another provider's model gets a clear 400, never a silent detour |
So if you'll use embeddings or files alongside chat, pick the format of the provider that serves them. The full list is on Endpoints.
When the format matters¶
- Provider-only features travel imperfectly. Anthropic's
cache_controlmarkers are removed on other providers (cache savings are still reported); thinking is mapped to the other side's reasoning setting, without Anthropic's signatures; Anthropic'sbashandtext_editorserver tools are refused on other providers. Gemini's own tools (such as Google Search) are Gemini's. What translation changed is listed in thex-vidai-adaptedheader. - A parameter with no equivalent is dropped, clamped or approximated, never a failed call, and named in
x-vidai-adapted. - Libraries that default to the Responses API (some versions of the OpenAI Agents SDK, the AI SDK, LangChain, DSPy and Pydantic AI) reach only OpenAI and Azure OpenAI models that way; each page says how to switch them to Chat Completions.
Pages by SDK¶
- OpenAI SDK for Python · for TypeScript
- Anthropic SDK for Python · for TypeScript
- Google GenAI SDK for Python · for TypeScript
- AI SDK by Vercel
- Frameworks and tools built on them: all integrations
Good to know¶
- An organisation can set a key to keep near-equivalent settings when translating (the preserve profile); a call
can ask for it with
x-vidai-translation: preservewhere the key allows it. - Translation adds a millisecond or two; the model's own time dominates.